The EU AI Act passed, but Mistral parties on | The AI Beat

Following a three-day marathon of negotiations, the EU AI Act finally passed over the weekend with Carme Artigas, the Spanish secretary of state for digitalization and artificial intelligence, calling it “a historical achievement, and a huge milestone towards the future!”

Among other agreements, the EU AI Act requires systems labeled as ‘high-impact general purpose AI models’ to adhere to transparency standards, while those classified as ‘high risk’ are subject to extra requirements including risk management, monitoring of severe incidents, evaluating the AI model, and implementing red teaming strategies.

In other words, the EU AI Act was massive, champagne-popping AI news — for about a minute. It lost some media momentum when it became clear that the Act is a provisional agreement, so it could take months before the bill gets its final wording. In addition, it won’t be enforced anytime soon — the AI Act won’t take effect until two years after final approval from European lawmakers. Um…two years? That’s the equivalent of two centuries in the AI hype cycle.

The EU AI Act was also quickly pushed from the headlines by one of Europe’s own AI stars. Paris-based open source model startup Mistral, in the equivalent of a mic drop, released a new LLM on Friday with nothing but a torrent link. Then, yesterday the company, which was founded just seven months ago by Meta and Google researchers, kept the party going by announcing a $415 million fund-raise from investors including Andreessen Horowitz, Lightspeed Venture Partners, Nvidia and Salesforce — valuing the small 22-person team at about $2 billion.

Mistral had led the charge against regulating foundation models

It’s notable that Mistral had been at the forefront of the EU AI Act negotiations a month ago — it lobbied against the European Parliament’s proposal for a tiered approach to regulating generative AI that ensured foundation models had guardrails. And many pointed to the fact that Mistral had deep government connections — Cedric O, France’s former state secretary for digital, was the one who led Mistral’s EU AI Act lobbying efforts. 

As recently as October, O had said in an interview that the EU AI Act could be “a zero or one regulation” for European generative AI startups like Mistral. “Depending on the [final shape of the] AI Act, it could kill Mistral, it could allow Mistral to grow, and there is a lot of gray zone in between,” had said. “It’s really critical.” 

In the end, Mistral seems to have succeeded in its lobbying efforts: The agreed-upon EU AI Act gives broad exemptions to open source models (though they do not apply to open source models deemed to pose a systemic risk). And its big wins over the past few days make clear that the EU AI Act is certainly not the end of the story when it comes to AI regulation in the European Union.

An article in TechCrunch made a good point about the debate going on there right now: Union: “Mistral’s good fortune is in many ways a microcosm of the fight for AI sovereignty. The European Union (EU) desires to avoid being left behind in yet another technological leap while at the same time imposing regulations to guide the tech’s development. As Germany’s Vice Chancellor and Minister for Economic Affairs Robert Habeck was recently quoted as saying: “The thought of having our own sovereignty in the AI sector is extremely important. [But] if Europe has the best regulation but no European companies, we haven’t won much.”

AI regulation will continue to play catch-up

No matter what, it’s clear that AI regulation — whether it is in the EU, in the US, or elsewhere — will be playing catch-up for a long time to come. The word that Mistral’s new Mixtral 8x7B model eclipsed the performance of OpenAI’s GPT 3.5 came as thousands of AI researchers, including those from all of the top AI labs, gathered at NeurIPS (Conference and Workshop on Neural Information Processing Systems) in New Orleans.

Those researchers, fresh off a mind-bending year in generative AI, are also ready to party — once they’ve finished presenting their papers, that is. They will raise their glasses in a tech toast to even faster AI development in 2024, while policy makers around the world try to figure out how to rein it in without turning the dial down on innovation. If Mistral’s high-flying celebration is any indication, regulators have one dire need right now. A need for speed.